Model training method and device for false information detection and positioning

Abstract

The invention relates to a model training method and device for false information detection and positioning in the field of information processing, and the method comprises the steps: carrying out the marking of each image-text sample in turns through a preset marking model and a prompt template, generating a structured marking result, and obtaining a training sample based on the marking result, the training sample comprises a label sample and a multi-round chain type thinking CoT sample; training a false information detection and positioning model based on a label sample and a CoT sample, comprising: a first stage: training a multi-modal model based on the label sample to obtain an initial model; in the second stage, label supervision training is conducted on the initial model based on CoT samples, a false information detection and positioning model is obtained, standard multi-round dialogues serve as context input in a teacher forcing mode, and the model learns the corresponding relation between the middle observation description and the structured label in each round of output. According to the invention, the false information detection and positioning precision is improved.

Description

Technical Field

[0001] This application belongs to the field of information processing technology, and more specifically, relates to a model training method and device for detecting and locating false information. Background Technology

[0002] With the rapid development of generative models and large language models, misinformation is rampant online, evolving from single-image forgery to cross-modal forgery, such as text forgery, audio-visual forgery, and combinations thereof, seriously threatening information integrity and user privacy. In sensitive areas such as politics and public health, facial recognition and tampering-related misinformation, by forging the facial identities and expressions of public figures or altering text descriptions, not only amplifies misleading and deceptive effects but may also cause reputational damage, manipulation of public opinion, and irreversible negative impacts on public safety and social stability.

[0003] Therefore, the needs in real-world scenarios have gradually evolved from simple binary classification for authenticity identification to tamper type recognition, and further refined into fine-grained tasks such as modal tampering region localization. Summary of the Invention

[0004] The purpose of this application is to provide a model training method and device for detecting and locating fake information, so as to improve the accuracy of fake information detection and location.

[0005] A first aspect of this application provides a model training method for detecting and locating false information, comprising: Using a pre-defined annotation model and prompt template, each image and text sample is annotated in rounds to generate structured annotation results. Training samples are obtained based on the annotation results. The training samples include: label samples and multi-round chained thinking CoT samples; label samples are samples containing user requests and the final structured annotation results; CoT samples are samples containing complete instructions and multi-round interactive annotation results. A fake information detection and localization model is trained based on labeled samples and CoT samples, including: Phase 1: Train the multimodal model based on labeled samples to obtain the initial model; The second stage involves supervising the initial model with labels based on CoT samples to obtain a false information detection and localization model. In this stage, standard multi-turn dialogues are used as context inputs through teacher-mandated methods, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each round of output.

[0006] In one embodiment, generating structured annotation results for each image / text sample through round-by-round annotation includes: First round: For each input image and text sample, label it according to the first set of observation dimensions, and generate labeling results for image authenticity and coarse localization; Second round: For each input image and text sample, label it with a preset second set of observation dimensions to generate labeling results of face manipulation type and fine-grained positioning; The third round: For each input image and text sample, annotations are performed under the image semantic constraints using a preset third set of observation dimensions to generate annotation results for text tampering type and token-level location; Fourth round: Output the final structured annotation result for each input image and text sample.

[0007] In one embodiment, the structured annotation results include: real samples and fake samples; obtaining training samples based on the annotation results includes: For fake samples, only the annotation results corresponding to each observation dimension in each round that are fake and whose number is greater than the first threshold, as well as the first type of image and text samples corresponding to the annotation results, are retained as training samples, and the structured annotation results of the first type of image and text samples are consistent with the real labels. For real samples, only the annotation results corresponding to each observation dimension in each round that are real and whose number is greater than the second threshold, along with the corresponding second type of image and text samples, are retained as training samples, and the structured annotation results of the second type of image and text samples are consistent with the real labels.

[0008] In one embodiment, the first stage includes: Label-supervised pre-training of multimodal models based on labeled samples enables the initial model to have basic false information detection and localization capabilities.

[0009] In one embodiment, standard multi-turn dialogue is used as contextual input through teacher-mandated methods, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each turn's output, including: For the t-th training round, the annotation results of the previous t-1 rounds are injected as context, and the loss is calculated only for the response of the t-th round; The total loss is obtained by accumulating the losses from each round, and the training is judged to end based on the total loss.

[0010] In one embodiment, the method further includes: Phase 3: Optimize the false information detection and localization model using reward evaluation and reinforcement learning.

[0011] In one embodiment, the third stage includes: For a complete multi-round inference trajectory, the total reward is calculated based on the observation dimensions of the multiple rounds and the weights corresponding to each observation dimension. The GRPO reinforcement learning algorithm is used to update and optimize the fake information detection and localization model based on the total reward.

[0012] In one embodiment, before generating structured annotation results by annotating each image and text sample in rounds using a preset annotation model and prompt template, the method further includes: preprocessing the initial image and text samples to obtain image and text samples with consistent format and alignment; including: For the image portion in the initial image and text samples, a fill-based preprocessing strategy is adopted to adjust the image size to a uniform size without changing the original pixel geometry. For the text portion in the initial image and text sample, perform sentence segmentation, word segmentation, or sub-word segmentation, and establish a mapping relationship between the token and the original character position.

[0013] A second aspect of this application provides a method for detecting and locating false information, including: Obtain the information to be tested; The false information detection and localization model trained using any of the methods in the claims is used to detect and localize the information to be detected, and to obtain the detection result of whether the information is false information.

[0014] A third aspect of the embodiments of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any model training method for false information detection and localization of this application, or the steps of any false information detection and localization method of this application.

[0015] The model training method and device for detecting and locating false information provided in this application have at least the following effects: This application improves the accuracy of the model's overall detection and localization by explicitly modeling the dependencies between subtasks through reconstructing the task into a multi-round chain-like reasoning process during model training.

[0016] In addition, during model training, subsequent rounds of reasoning are performed under the conditional constraints of the previous round's results, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each round's output. The model possesses a multi-round thinking path "from overall judgment to local location and then to text tracing," which improves the model's ability to comprehensively utilize complex multimodal tampering clues. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1A schematic flowchart of a model training method for detecting and locating false information provided in an embodiment of this application; Figure 2 This is a schematic diagram of a multi-round chain-like thought annotation process according to an embodiment of this application; Figure 3 This is a model training architecture diagram of an embodiment of this application; Figure 4 A flowchart illustrating a method for detecting and locating false information provided in an embodiment of this application; Figure 5 This is a schematic diagram of an electronic device structure according to an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of this application.

[0020] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Where there is no conflict between the embodiments, the following embodiments and features can be combined with each other. The steps in the following method embodiments are for illustrative purposes only and are not intended to limit this application.

[0021] Existing methods for detecting and locating misinformation have at least the following problems: 1) Fragmented multi-task operation makes it difficult to model cross-task relationships. Existing multimodal detection models typically adopt a structure of "one backbone network + multiple task heads," outputting results separately for sub-tasks such as authenticity determination, image tampering type recognition, image tampering region localization, and text tampering detection. The lack of explicit dependency modeling and information transfer between tasks leads to inconsistencies in conclusions between different sub-tasks; for example, the overall result may be true, but some sub-tasks may still output tampering bounding boxes.

[0022] 2) Lack of process-level reasoning supervision leads to insufficient interpretability. Most existing methods only supervise the final label—true / false, type, or location—without constraining the intermediate reasoning process during training. In complex scenarios, the model may obtain the correct label through speculation, but the internal reasoning logic may be flawed or unstable. The output often only contains a final category or bounding box, lacking explanation of the process based on which visual or semantic evidence led to the judgment. This makes it difficult to meet the high requirements of interpretability and traceability for content security review and legal evidence collection. Consequently, the system is prone to performance fluctuations when facing distribution shifts or novel forged samples, resulting in poor overall robustness.

[0023] refer to Figure 1 , Figure 1 This is a flowchart illustrating a model training method for detecting and locating fake information according to an embodiment of this application. The method may include: S11: Use the preset annotation model and prompt template to annotate each image and text sample in rounds to generate structured annotation results. Based on the annotation results, obtain training samples, which include: label samples and CoT (multi-round chain thinking) samples.

[0024] The preset annotation model and prompt template can be any existing multimodal annotation model that can automatically annotate text; this application does not impose any specific restrictions on them.

[0025] In one embodiment, the label sample is a sample containing user requests and the final structured annotation result; the CoT sample is a sample containing complete instructions and multi-round interactive annotation results, that is, it not only contains user requests and the final structured annotation result, but also contains multi-round interactive instructions and intermediate annotation results throughout the process.

[0026] In one embodiment, a schematic diagram illustrating the generation of structured annotation results for each image and text sample through round-by-round annotation using a preset annotation model and prompt template is shown below. Figure 2 As shown, the image and text samples and prompts are input into the annotation model, and then multiple rounds of annotation are performed. The multiple rounds of annotation specifically include: First round: For each input image and text sample, label it according to the first set of observation dimensions, and generate labeling results for image authenticity and coarse localization; Second round: For each input image and text sample, label it with a preset second set of observation dimensions to generate labeling results of face manipulation type and fine-grained positioning; The third round: For each input image and text sample, annotations are performed under the image semantic constraints using a preset third set of observation dimensions to generate annotation results for text tampering type and token-level location; It should be noted that this application sets a set of observation dimensions for each round, and the observation dimensions for each round can be different. These observation dimensions may include, for example, structural consistency, lighting consistency, character identity consistency, and behavioral semantic consistency.

[0027] The fourth round: summarizing the output, that is, outputting the final structured annotation results for each input image and text sample.

[0028] In one embodiment, the structured annotation results include: real samples and fake samples; real samples are text and image samples that have not been tampered with or forged; fake samples are text and image samples that have been tampered with or forged.

[0029] After multiple rounds of annotation, multiple rounds of annotation results are generated. This application can perform rule-based filtering and consistency verification on the generated annotation results. The rule-based filtering process involves obtaining training samples based on the annotation results, including: For fake samples, only the annotation results corresponding to the observation dimensions in each round that are fake and whose number is greater than the first threshold, as well as the first type of image and text samples corresponding to the annotation results, are retained as training samples. For real samples, only the annotation results corresponding to the observation dimensions of each group of observations in each round that are considered real and whose number is greater than the second threshold, as well as the corresponding second type of image and text samples, are retained as training samples.

[0030] Among them, consistency verification ensures that the annotation results of the first type of image and text samples selected are consistent with the real labels, and the annotation results of the second type of image and text samples are consistent with the real labels.

[0031] In one embodiment, for the acquired training samples, the label names of each dimension may not be retained, but the annotation results of each dimension may be retained.

[0032] This application utilizes a labeling model to automate the labeling of sample data and completes multi-round reasoning, evidence generation, and result output within a unified model framework, avoiding the structural complexity and maintenance costs associated with cascading multiple models. Simultaneously, the labeling model assists in generating fine-grained, multi-round reasoning labels, and automatically constructs training datasets by combining rule-based filtering and consistency verification, significantly reducing the workload of manual labeling and system deployment costs, and achieving efficient acquisition of high-quality, large-scale, multi-round thought chain data.

[0033] Furthermore, traditional methods only label the final tags, neglecting the intermediate reasoning process. This results in the model only providing a "conclusion" without explaining "how the conclusion was reached," hindering manual review and judicial evidence collection. This application addresses this by designing intermediate dimensions (observation dimensions corresponding to each round) and introducing multiple verifications to ensure consistency between the labeled results and the actual labels. This constructs high-quality, multi-round CoT training sample data, allowing the reasoning chain to be output in a natural language and structured format. This significantly improves the system's interpretability and auditability, making it particularly suitable for business scenarios requiring "reasonable and well-founded explanations."

[0034] In one embodiment, before generating structured annotation results by annotating each image and text sample in rounds using the annotation model and prompt template, the method further includes: preprocessing the initial image and text samples to obtain image and text samples with consistent format and alignment; including: For the image portion in the initial image and text samples, a fill-based preprocessing strategy is adopted to adjust the image size to a uniform size without changing the original pixel geometry. For the text portion in the initial image and text sample, perform sentence segmentation, word segmentation, or sub-word segmentation, and establish a mapping relationship between the token and the original character position.

[0035] To adapt to the input size of deep networks, existing systems generally employ preprocessing strategies such as scaling. This alters the geometric proportions and local texture distribution of the original image, destroying subtle tampering traces such as facial edges and blending gaps. This application employs a fill-based preprocessing strategy for the image portion. Without changing the original pixel geometry, it adjusts the image to a uniform size suitable for the input of a large multimodal model, and synchronously translates and aligns existing bounding box annotations to ensure a one-to-one correspondence between the annotation boxes and pixel positions. Simultaneously, the text portion is segmented into sentences, words, or sub-words, establishing a mapping relationship between tokens and the original character positions. The image-text samples after this preprocessing possess a unified format and precisely aligned multimodal structure, providing a standardized input foundation for subsequent chain-like annotation and model training in this application.

[0036] S12: Training a fake information detection and localization model based on labeled samples and CoT samples.

[0037] In one embodiment, such as Figure 3 As shown, model training includes at least the following two stages: Phase 1: Train the multimodal model based on labeled samples to obtain the initial model; The first stage involves label-supervised pre-training of the multimodal model based on labeled samples, enabling the initial model to possess basic false information detection and localization capabilities.

[0038] In other words, the first stage involves training the multimodal model using the labeled samples obtained in step S11. This enables the multimodal model to accurately determine the true / false status, image manipulation type, image bounding box, and the location of text tampering words in a single output round, thus aligning with the task format and basic decision boundaries. This stage aims to establish an initial model with basic detection and localization capabilities.

[0039] The second stage involves supervising the initial model with labels based on CoT samples to obtain a false information detection and localization model. In this stage, standard multi-turn dialogues are used as context inputs through teacher-mandated methods, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each round of output.

[0040] This stage further constrains the logical consistency and semantic coherence of the reasoning chain itself, beyond label supervision.

[0041] In one embodiment, standard multi-turn dialogue is used as contextual input through teacher-mandated methods, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each turn's output, including: For the t-th training round, the annotation results of the previous t-1 rounds are injected as context, and the loss is calculated only for the response of the t-th round; The total loss is obtained by accumulating the losses from each round, and the training is judged based on the total loss. This allows the model to gradually learn a multi-round thinking path, from overall judgment to local localization and then to text source tracing.

[0042] The total loss can be expressed by the following formula (1).

[0043]

[0044] As shown in formula (1): This represents the input training samples, such as image tensors, text token sequences, etc. This represents the complete sequence of multi-round CoT standard answers (i.e., the labeled results corresponding to each round), where the first... wheel It also includes intermediate observation descriptions for that round (i.e., the annotation results corresponding to each round) and structured annotation results; Indicates the first The first round of output sequence One token, This is the length of the round sequence; Indicates the preceding The "standard answer" for the first round of training The context is injected through a teacher-mandated approach during turn-based instruction. Indicates the previous signal provided by the teacher in the current round. One token; For parameters The conditional probabilities given by the multimodal model; This is the CoT training sample dataset; Total loss That is, to accumulate the total loss of all rounds at the sample level, so that the model "is able to achieve the desired result in a given timeframe". Learning the first round of standard answers The correct output of the round is achieved during the training. When the wheel is output, the front The standard answer of each round is injected as context. The loss is calculated only for the response of the current round, and the loss of each round is accumulated at the sample level, so that the model gradually learns a multi-round thinking path from overall judgment to local localization and then to text tracing.

[0045] In existing technologies, subtasks are connected in parallel to the same backbone network, lacking explicit dependencies, which easily leads to contradictions between truth / falseness determination and local localization, and between image conclusions and text conclusions. This application reconstructs the task into a multi-round chain-like reasoning process during model training, allowing image authenticity determination, face manipulation type recognition, fine image localization, and text token-level localization to unfold progressively "from coarse to fine, from image to text." Subsequent rounds reason under the constraints of the previous round's results, improving the model's ability to comprehensively utilize complex multimodal tampering clues.

[0046] In one embodiment, model training further includes: Phase 3: Optimize the false information detection and localization model using reward evaluation and reinforcement learning.

[0047] In one embodiment, the third stage includes: For a complete multi-round inference trajectory, the total reward is calculated based on the observation dimensions of the multiple rounds and the weights corresponding to each observation dimension. The false information detection and localization model is updated and optimized based on the total reward using the GRPO (Group-wise Relative Policy Optimization) algorithm.

[0048] Specifically, a multi-dimensional reward function system is first constructed for multi-round inference outputs. For each complete multi-round inference trajectory... —This includes the full output from the input training samples to the final structured conclusion and the intermediate multiple rounds of reasoning. Sub-rewards are calculated based on multiple observation dimensions, such as information authenticity, tampering type, image localization, text localization, and output format. The total reward is obtained through weighted fusion, as shown in the formula below. :

[0049] As shown in formula (2): in, The non-negative weights for each observation dimension can be set or adjusted according to specific business needs; in, This represents the information-level true / false reward: a positive reward is given when the model's predicted true / false conclusion matches the true label, and a zero reward is given when they do not match. To reward the image manipulation type, the model's predicted image manipulation type is compared with the true label. Different segmented rewards are assigned to complete match, partial match and no match, as shown in formula (3).

[0050] Image localization reward; bounding boxes are output based on the model. With the true bounding box The IoU values ​​between the points are used to calculate continuous or segmented rewards, as shown in formula (4);

[0051] For text tampering type rewards, the model outputs text tampering types are compared with the real labels, and a segmented reward design is also adopted; To achieve fine-grained reward localization for text, the token mapping established during the preprocessing stage is used to map the set of tampered tokens predicted by the model. With the real set In comparison, the F1 value is calculated and used as a reward, as in formula (5);

[0052] To ensure structured output and consistency rewards, check whether the fields of the outputs from multiple rounds are complete, whether the format is correct, and whether the conclusions between rounds are consistent. If all constraints are met, a reward is given; otherwise, zero reward is given. After receiving the above total reward Subsequently, a GRPO-based optimization strategy was used to update the multimodal large model.

[0053] Specifically, multiple candidate inference trajectories generated under the same input or the same batch of tasks are divided into several groups, denoted as the first group. The set of trajectories in the group is The corresponding total reward is Within each group, first calculate the average reward for that group. and standard deviation For example, in formula (6):

[0054] The reward for each trajectory is then normalized within the group to obtain a relative advantage value, which is then used for subsequent iterative optimization. During training, the optimization objective is to minimize the GRPO (Gross Response Point Objective). This is equivalent to following the relative advantage within the group. In the direction of the following formula (7), the generation probability of the trajectory with higher reward is increased and the trajectory with lower reward is suppressed. Thus, under the premise of maintaining the multi-round inference structure and process supervision constraints, the model can achieve an overall improvement in key business indicators such as information authenticity, modification type, image IoU, text token F1 and output format legality.

[0055]

[0056] The third stage, building upon the first two stages, combines multidimensional rewards and GRPO strategy optimization methods to perform joint reinforcement learning on the model at both the process and outcome levels, further improving its performance on business metrics such as news authenticity, tampering type, IoU, token-F1, and the legality of structured output.

[0057] The model training method for fake information detection and localization in this application unifies multimodal fake information detection and manipulation localization into a multi-round chain-like reasoning process. This allows authenticity determination, image tampering type recognition, fine image localization, text tampering detection, and token-level localization to be executed sequentially and mutually constrained within the same large multimodal model. This explicitly models the dependencies between subtasks, thereby improving the overall detection and localization accuracy.

[0058] This application employs a three-stage training paradigm: "first-stage label-supervised pre-training—second-stage joint label and CoT supervision—third-stage multi-dimensional reward RL optimization," progressively breaking down the difficulty. The first stage ensures the model learns to stably output structured labels, addressing the "can it solve problems" issue. The second stage introduces multiple rounds of CoT on top of the existing foundation, enabling the model to learn "how to reason step by step." The third stage aligns key business metrics with RL and multi-dimensional rewards. Because the input, objective, and loss function of each stage are relatively clear, and the parameters are initialized using the previous stage, the training process converges more smoothly, significantly reducing the oscillation and divergence risks of direct RL training. This ensures the model possesses complex reasoning capabilities while maintaining good training stability and output quality.

[0059] See Figure 4 The flowchart below illustrates a method for detecting and locating false information provided in this application. The method includes: S41. Obtain the information to be detected.

[0060] This application can be used to detect and locate the authenticity of various types of information. The information to be detected will vary depending on the specific scenario.

[0061] In one embodiment, in news media and various content platforms, this application can be used to automatically verify the authenticity of news articles and their accompanying images, identify and mark image areas and text fragments that may be tampered with, and generate a complete multi-round reasoning evidence chain.

[0062] In one embodiment, in the scenario of monitoring harmful content and misinformation on social media, this application can be used to perform batch detection of large-scale multimodal posts published by users, automatically identify complex manipulation behaviors such as face replacement and malicious alteration of captions, and provide corresponding risk scores and explainable prompts to assist platform operators in carrying out content security governance, user behavior compliance management, and public opinion risk warning.

[0063] In one embodiment, in the context of judicial evidence collection and electronic evidence identification, this application is used to conduct technical identification of the graphic materials involved in the case, automatically determine whether the relevant content has been tampered with, the specific method of tampering and the precise location, and output corresponding multi-round reasoning explanations and evidence chains, thereby enhancing the technical support and reviewability of electronic evidence and meeting the requirements of judicial institutions for the objectivity and interpretability of evidence.

[0064] In one embodiment, in fields such as finance, government affairs, and public relations where the authenticity of information is highly sensitive, this application can be used to quickly screen graphic announcements, public statements, and online public opinion content involving major matters, identify potential forged or misleading information, help decision-making departments to promptly identify risk points and take intervention measures, and reduce the impact of erroneous information on the public and the market.

[0065] S42 uses a false information detection and localization model to detect and localize the information to be detected, and obtains the detection result of whether the information is false information.

[0066] The fake information detection and localization model is a model trained by the method of any of the above embodiments.

[0067] See Figure 5 , Figure 5 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 5 The electronic device 500 shown in this embodiment may include one or more processors 501, one or more input devices 502, one or more output devices 503, and one or more memories 504. The processors 501, input devices 502, output devices 503, and memories 504 communicate with each other via a communication bus 505. The memories 504 store computer programs, including program instructions. The processors 501 execute the program instructions stored in the memories 504.

[0068] It should be understood that, in the embodiments of this application, the processor 501 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0069] Input device 502 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 503 may include a display (LCD, etc.), a speaker, etc.

[0070] The memory 504 may include read-only memory and random access memory, and provides instructions and data to the processor 501. A portion of the memory 504 may also include non-volatile random access memory.

[0071] In specific implementations, the processor 501, input device 502, and output device 503 described in the embodiments of this application can execute the implementation method described in the model training method for false information detection and localization provided in the embodiments of this application, or they can execute the implementation method of the electronic device described in the embodiments of this application, which will not be repeated here.

[0072] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0073] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., provided on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0074] This application provides a computer program product, which includes computer-executable instructions or a computer program. The computer-executable instructions or computer program are stored in a computer-readable storage medium. The processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the implementation described in the model training method for false information detection and localization provided in this application, or the implementation described in the false information detection and localization method.

[0075] Those skilled in the art will recognize that the modules / units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules, units, or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or modules / units, or it may be an electrical, mechanical, or other form of connection.

[0078] The modules / units described as separate components may or may not be physically separate. Similarly, the components shown as modules / units may or may not be physical modules / units; they may be located in one place or distributed across multiple network modules / units. Some or all of the modules / units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0079] Furthermore, the functional modules / units in the various embodiments of this application can be integrated into one processing module / unit, or each module / unit can exist physically separately, or two or more modules / units can be integrated into one module / unit. The integrated modules / units described above can be implemented in hardware or in the form of software functional modules / units.

[0080] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A model training method for detecting and locating false information, characterized in that, include: The system uses a preset annotation model and prompt template to annotate each image and text sample in rounds to generate structured annotation results. Based on the annotation results, training samples are obtained. The training samples include: label samples and multi-round chained thinking CoT samples. The label samples are samples containing user requests and the final structured annotation results. The CoT samples are samples containing complete instructions and multi-round interactive annotation results. Training a fake information detection and localization model based on the labeled samples and CoT samples includes: Phase 1: Train the multimodal model based on the labeled samples to obtain the initial model; The second stage involves supervising the initial model with labels based on the CoT samples to obtain the false information detection and localization model. In this stage, standard multi-turn dialogues are used as context inputs through teacher-mandated methods, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each round of output.

2. The method as described in claim 1, characterized in that, The process of generating structured annotation results for each image and text sample through round-by-round annotation includes: First round: For each input image and text sample, label it according to the first set of observation dimensions, and generate labeling results for image authenticity and coarse localization; Second round: For each input image and text sample, label it with a preset second set of observation dimensions to generate labeling results of face manipulation type and fine-grained positioning; The third round: For each input image and text sample, annotations are performed under the image semantic constraints using a preset third set of observation dimensions to generate annotation results for text tampering type and token-level location; Fourth round: Output the final structured annotation result for each input image and text sample.

3. The method as described in claim 2, characterized in that, The structured annotation results include: real samples and fake samples; obtaining training samples based on the annotation results includes: For fake samples, only the annotation results corresponding to each observation dimension in each round that are fake and whose number is greater than the first threshold, as well as the first type of image and text samples corresponding to the annotation results, are retained as training samples, and the structured annotation results of the first type of image and text samples are consistent with the real labels. For real samples, only the annotation results corresponding to each observation dimension in each round that are real and whose number is greater than the second threshold, along with the corresponding second type of image and text samples, are retained as training samples, and the structured annotation results of the second type of image and text samples are consistent with the real labels.

4. The method as described in claim 1, characterized in that, The first stage includes: Based on the labeled samples, the multimodal model is pre-trained with label supervision, so that the obtained initial model has basic false information detection and localization capabilities.

5. The method as described in claim 1, characterized in that, The method of using standard multi-turn dialogue as contextual input through teacher-mandated instruction, enabling the model to learn the correspondence between intermediate observation descriptions and structured labels in each turn's output, includes: For the t-th training round, the annotation results of the previous t-1 rounds are injected as context, and the loss is calculated only for the response of the t-th round; The total loss is obtained by accumulating the losses from each round, and the training is then judged to be complete based on the total loss.

6. The method as described in claim 1, characterized in that, The method further includes: The third stage involves optimizing the false information detection and localization model using reward evaluation and reinforcement learning.

7. The method as described in claim 6, characterized in that, The third stage includes: For a complete multi-round inference trajectory, the total reward is calculated based on the observation dimensions of the multiple rounds and the weights corresponding to each observation dimension. Based on the total reward, the false information detection and localization model is updated and optimized using the reinforcement learning algorithm GRPO.

8. The method as described in claim 1, characterized in that, Before generating structured annotation results by annotating each image and text sample in rounds using a preset annotation model and prompt template, the method further includes: preprocessing the initial image and text samples to obtain image and text samples with consistent format and alignment; including: For the image portion in the initial image and text sample, a fill-based preprocessing strategy is adopted to adjust the image size to a uniform size without changing the original pixel geometry. For the text portion in the initial image and text sample, perform sentence segmentation, word segmentation, or sub-word segmentation to establish a mapping relationship between the token and the original character position.

9. A method for detecting and locating false information, characterized in that, include: Obtain the information to be tested; The false information detection and localization model trained by any one of claims 1 to 8 is used to detect and localize the information to be detected, and the detection result of whether the information is false information is obtained.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as claimed in any one of claims 1 to 8, or the steps of the method as claimed in claim 9.

Images